Liquidity sweeps — does price reverse after one?

37,162 sweeps·Seven futures markets, 2019–2026·Sep 2026

Every sweep of a previous-day, Asia, London or prior-week high or low in seven futures markets, 2019 to mid-2026 — 23,718 of them — against a fake level matched to each one and run through the same code. Built on 2019–2024, then checked on 2025–2026. In all, 37,162 sweeps across the 14 level types tested; the four most traded are shown here.

  • 49.9%After a sweep, price reverses 49.9% of the time. After a sweep of a matched fake level, 50.0%. A coin flip — and 0 of 97 markets and levels did better.
  • 1 / 263confirmation combinations made money at a one-to-one target on real levels. On the fake levels, up to 28 did.
  • +5.9 ptsThe one real effect runs the other way. At the first touch, real previous-day highs and lows get run through 56.0% of the time; fake levels, 50.1%.
  • 8.4×Volume in the first second at a real level, against 5.5× at a fake one. The stops really are there — and they are gone in a second.
The pattern

What a liquidity sweep is

A liquidity sweep is when price takes out a key level and then closes back inside it. The idea is that stop orders rest just beyond obvious highs and lows; price runs them, the stops get filled, and with nobody left to push it further, price turns around. We tested it on the levels most traders mark: the previous day’s high and low, the Asia and London session highs and lows, and the prior week’s.

Level · seven markets, 2019–2026SweepsFirst touches
Previous day high / low6,00913,587
Asia session high / low8,54619,087
London session high / low8,12017,768
Prior week high / low1,0432,460
All four23,71852,902

Markets: NQ, ES, YM, GC, SI, CL and RTY. Sweeps found on five-minute candles; outcomes measured on one-minute bars.

The method

What each sweep is compared with

The fake level. To know whether a level matters, you need something to compare it with. So for every real level we built a fake one — a price that isn’t a key level, but matches it in every other way:

  • Same time of day, on a different day
  • Same distance from price, adjusted for how fast the market was moving
  • Similar volatility that day
  • Same side of price, above or below

We ran the exact same code on both. If real levels matter, they have to beat the fakes.

The bar. Every test was written down before we looked at the data. To count, a result had to:

  • Hold on the data the study was built on (2019–2024) and on data it wasn’t (2025–2026)
  • Survive a correction for running many tests — test enough things and some look good by luck
  • Beat what the same test finds on randomly shuffled prices

One disclosure: some of the 2025–2026 data had been looked at by earlier studies of these levels, so that check is a second look rather than a first one.

How the code was proved

A bug that finds nothing looks exactly like a pattern that predicts nothing, so the measuring equipment was checked before any answer was trusted. Every level value matched an independent rebuild — 110,504 of them — and every touch, sweep and break matched a minute-by-minute brute-force search.

The shuffled-price world. We shuffled each day’s one-minute price moves into a random order — the same raw ingredients, with no real buying and selling left in them — and ran everything again. A few thin cells, mostly weekly and monthly levels, still showed differences of 2 points or more there. Each cell’s shuffled-world difference is subtracted before it can count.

The words used on this page, in plain English

ATR — how far the market typically moves in one five-minute candle. Using it instead of dollars puts the Nasdaq and gold on the same scale.

A point — one percentage point. If real levels reverse 51% of the time and fake ones 50%, the real level is one point ahead.

R — the amount risked on a trade. “−0.20R per trade” means the average trade lost a fifth of what it risked.

In sample / out of sample — the study was built on 2019–2024. 2025–2026 was used only to check the result.

No difference / too few to tell — “no difference” means there was enough data to rule out an edge of about one to two points. “Too few to tell” means there were not enough events to say either way.

Costs — commission plus one tick of slippage on the entry and on every exit. A tick is the smallest step a market’s price moves in.

Topic 1 · The reversal

Does price reverse after a sweep?

From the close of the sweep candle we set a target one ATR away in the reversal direction and a stop one ATR back through the level. Does price reach the target before the stop? Both are checked only from the next minute on, and a tie counts as a loss. Fifty per cent is a coin flip before costs.

Reverses one ATR before going one ATR back throughAfter a real sweepAfter a fake sweepDifference
All four level groups49.9%50.0%−0.1
Previous day high / low50.4%49.6%+0.8
Asia session high / low50.4%50.5%−0.1
London session high / low49.3%50.1%−0.8
Prior week high / low49.0%48.8%+0.2

A coin flip. Across all 23,718 sweeps the difference is −0.1 points, with a 95% range of −0.9 to +0.9 — enough data to have spotted an edge under one point. A measured zero, not a shortage of data.

Every market on its own, in and out of sample

In case one market was hiding an effect that the pooled number averages away, every market was run on its own, for every level group and both sweep definitions — 97 comparisons. None survived the correction for the number of tests, and none was ahead in both windows. Below is every high and low pooled, on the five-minute sweep.

MarketReal · in sampleFake · in sampleDifference · in sampleDifference · out of sample
All seven49.7%49.9%−0.2+0.4
NQ · Nasdaq 10049.9%50.5%−0.6+2.4
ES · S&P 50048.6%49.8%−1.2+1.2
YM · Dow48.4%50.0%−1.7−1.9
RTY · Russell 200050.2%49.5%+0.7+0.7
GC · Gold50.9%50.0%+0.9+1.9
SI · Silver48.4%49.7%−1.3−1.1
CL · Crude oil51.3%49.9%+1.4−1.0

Every market sits within about two points of its fake level, and the sign flips between the two windows in half of them. 19,224 sweeps in sample and 4,730 out of sample across every high and low level type.

Topic 2 · The confirmations

Does anything make a sweep reverse?

Most traders say a sweep needs confirmation before you take it. So we tested every one that gets taught, trading each the same way:

  • Entry — the next one-minute candle after the confirmation
  • Stop — two ticks past the sweep’s wick
  • Target — the same distance as the stop

Then we ran the exact same trades at the fake levels.

Trading the reversal · target 1R, after costsReal · winsReal · per tradeFake · winsFake · per trade
No confirmation at all45.4%−0.23R45.2%−0.25R
With the trend43.6%−0.26R44.6%−0.26R
Break of structure45.2%−0.14R47.2%−0.11R
Fair value gap45.5%−0.14R49.1%−0.07R
Inverse fair value gap44.9%−0.15R48.2%−0.09R
Inside a killzone45.1%−0.23R44.3%−0.26R
On a volume surge44.1%−0.25R44.7%−0.25R
Clean, shallow sweep45.2%−0.26R45.4%−0.27R
Never tested before46.3%−0.21R45.5%−0.24R
Two levels stacked46.9%−0.20R46.5%−0.20R
Premium / discount45.1%−0.25R44.8%−0.27R

Every confirmation loses money on a real level.

  • The best ones — break of structure, a fair value gap and an inverse fair value gap lose less, but they do even better at fake levels
  • Stacking them — of 263 combinations of up to three, 1 made money on real levels
  • At fake levels — up to 28 did. That’s what testing hundreds of filters produces when there is nothing to find
The strictest version of the model

The most specific version taught says it isn’t any high or low that matters but a pool — equal highs or equal lows, where stops visibly cluster — swept with a decisive candle, in line with the daily bias, with a clear target on the other side. We built exactly that and tested it the same way: 8,269 real pool sweeps across the seven markets, against three separate sets of matched fake pools and a shuffled-price world.

Equal-high / equal-low pool sweeps · 2019–2024Result
Reverses after a real pool sweep vs a fake pool sweep50.6% vs 49.7%
A more decisive sweep candle · one ATR body or more−0.8 pts · worse, not better
Versions of the model that made money after costs0 / 54
Best version on real levels · per trade−0.20R
Best version in the shuffled-price world · per trade−0.005R

Each filter makes the losing trade slightly less of a loser; none makes it a winner. A world with no real buying and selling in it produced a better-looking setup than the real one.

Reading the order flow at the sweep

The last explanation left is that you have to read the tape: a sweep that gets absorbed — heavy volume that fails to move price — is the one that reverses. We measured every trade and the resting book at 69,602 touches across five markets, from one second to thirty seconds after the sweep.

It sorts sweeps in the wrong direction. The more a sweep is absorbed, the less it reverses — the most-absorbed fifth reverses 2.4 to 2.9 points less often than the least. No slice of any order-flow measure reaches 55%, the rate a one-to-one reversal needs; the best reaches 50.1%. The same sorting shows up at fake levels, and the size resting on the far side of the book at the moment of the sweep predicts nothing at all.

One measurement error was caught and fixed before any result was read: timing the outcome from the level rather than from the price at the end of the window made absorption look like it worked by 18 points. It was measuring its own definition.

Waiting for the reclaim, waiting longer, and entry tricks
VariationWhat happened
Wait for price to reclaim the level, then fadereal, and tiny — about 2 points over a fake level pooled (+0.02 to +0.05R before costs). Market by market, 17 of 352 looked good; chance alone predicts 17.6
Wait longer for the reversal97% of touches break the level and 96% of those come back — a failed break is the norm, not a signal. The longer the break lasts, the less it reverses: 41.7% within 15 minutes, 36.0% after four hours
Enter on a one-minute break of structureit follows 99.6% of sweeps, so it selects nothing — −0.32R per trade; the tighter stop it allows makes it worse, −0.47R
Enter on the retest of the swept level−0.83R per trade, 24% winners — the level does not hold

Of 298 versions of the last two entries, none made money after costs and none survived the correction for the number of tests. Waiting removes the loss; it never creates an edge.

Topic 3 · What does happen

The levels are real. Price runs through them.

If sweeps don’t reverse, do these levels mean anything at all? To find out, we stopped waiting for a sweep and looked at the first time price reaches each level. From there, does it break through or bounce back? We measured both the same distance — one ATR — and asked which came first, at real levels and at fake ones.

If the levels were just lines on a chart, real and fake would come out the same. They don’t — but not in the direction that gets taught.

Breaks one ATR through before bouncing one ATRReal levelFake levelDifference
Previous day high / low56.0%50.1%+5.9
Asia session high / low53.3%49.0%+4.3
London session high / low55.4%51.7%+3.7
Prior week high / low61.9%50.3%+11.6

Real levels get run through more, not less. 52,902 first touches, seven markets, 2019–2026. The difference holds in all seven markets, in every year, with and against the trend, at every approach speed, and keeps its sign out of sample. It is strongest at the most obvious extremes. The prior day’s and today’s 9:30 opens behave like any other price.

What the tape does in the first second at the level
Real level
8.4×
Fake level
5.5×

Trading volume in the first second at the level, compared with the 30 seconds before.

  • What it means — a burst that big is stop orders getting hit all at once. The stops are real.
  • Which way — 70.5% of that burst trades in the break direction. Price runs through the stops, not away from them.
  • Is it tradeable? — not as a simple bet. Betting on the break with an equal stop and target still loses after costs (−0.36R in sample, −0.22R out of sample). The edge is smaller than what each trade costs.
Every level, in and out of sample
How much more often a real level breaks than a fake one · pointsIn sample · 2019–24Out of sample · 2025–26
Previous day low+6.3+5.8
Previous day high+5.9+4.4
Asia high / low+4.1 / +5.1+2.5 / +3.7
London high / low+4.5 / +3.8+2.3 / +1.2
Prior week high / low+11.5 / +12.1+13.3 / +7.0
Prior day’s open · today’s open−1.1 / +1.0−1.1 / −1.2

Every high and low keeps its sign out of sample. The out-of-sample window holds about a quarter of the data, so it can only confirm an edge of about five points against about two and a half in sample; the previous-day low is the one level that clears the full bar in both. The two 9:30 opens show nothing either way. Real levels are also broken more than stale highs and lows from a week earlier: it is the current extremes that matter, not any old high.

The order flow in full, and the number that looks like a signal but isn’t
After price reaches the levelReal levelFake level
Volume, as a multiple of the 30 seconds before
The first second8.4×5.5×
1 to 5 seconds after1.4×1.1×
5 to 30 seconds after1.3×1.1×
Share of volume trading in the break direction
The 30 seconds before62.2%62.1%
The first second70.5%67.4%
1 to 5 seconds after50.2%49.9%

29,192 touches with every trade print, five markets, 2019–2024.

The 62% before the touch is not a signal. It looks like the market leaning into the level early — but a fake level reads 62.1% in the same window. To reach a price above you, price has to be going up; that number describes approaching any price, not this one. Only the first second separates real from fake, and by five seconds it is over.

Fresh levels against old ones

A common idea is that the longer a level survives untouched, the more orders pile up behind it. The data says the opposite. The run-through difference is largest the same day and fades away as the level ages; a level untouched for 20 or more trading days behaves like a random line.

Level untouched forRun-through vs fake · pointsFirst-second volume
Touched the same day+6.411.2×
1 day+5.19.0×
2 to 4 days+4.87.5×
5 to 9 days+2.67.0×
10 to 19 days+3.76.4×
20 days or more−0.25.3×

In sample, all level types pooled. The first-second burst at real levels shrinks steadily with age while fake levels barely move, so the gap between them collapses. Most levels are hit the same session they form.

From effect to strategy

Turning a real effect into a strategy

The effect was real, but trading it the simple way lost money. So I looked deeper into where the effect is strongest, how to capture it and how to manage the trade. Then I put the result through the same test as everything on this page: built on 2019–2024 and run once on data it had never seen, with costs and slippage included.

That became T-812. The rules stay private. The results don’t:

T-812TradesWin rateProfit factorSharpe
In sample · 2019–20241,29459.0%1.271.06
Out of sample · 2025–202631360.4%1.451.84

Hypothetical / backtested on tick data, $150,000 account risking 1% per trade. Sharpe is annualized from monthly returns, net of the risk-free rate. Past performance is not indicative of future results.

See T-812’s full track record →
Claimed vs measured

What is taught, next to what the data shows

What is taughtWhat 23,718 sweeps show
Price reversesafter a sweep
49.9% vs 50.0% — after a real sweep, and after a sweep of a matched fake level; 0 of 97 markets and levels
The stops get huntedthen price turns
56% vs 50% — the stops are real, but price runs through them; real previous-day levels break more often than fake ones
Add confirmationto find the good ones
1 of 263 — combinations profitable on real levels; up to 28 on fake levels
Untouched levelsbuild more liquidity
+6.4 → −0.2 pts — the effect is strongest the same day and gone after 20 days
Verdict

Liquidity sweeps don’t reverse. The levels are real — they get run through.

Price reverses after a sweep no more often than after a sweep of a matched fake level, in any market, at any level, with any confirmation we tested. At the first touch these levels get broken more often than empty price — the stops really are sitting there, and fading the level is trading the wrong side of them.

All seven predictions, written before the data was run
What we predictedWhat happened
1 · Price is drawn to these levels no more than to a fake levelright — 0 of 12
2 · Price reacts at the first touch no differently than at a fake levelwrong — real levels get broken more; the previous-day low passes the full bar in both windows
3 · Price reverses after a sweep no more than after a fake sweepright — 0 of 12
4 · A close beyond the level follows through no more than at a fake levelright — 0 of 12
5 · Price does not pause at these levelsright — 0 of 12
6 · Previous-day and weekly levels show a real reaction (in the concept’s favour)right — but as breaks, not bounces
7 · Today’s 9:30 open works as a pivot (in the concept’s favour)wrong — it behaves like any other price
The final audit. Every level and event on 25 random days of NQ and ES was rebuilt using only the data available at that minute — 370,068 decisions across 2,993 rebuilds, 0 mismatches. Three deliberately planted leaks were each caught first.
← Back to all studies

Now see what actually works.

The same brutal test — matched controls, real costs, one honest out-of-sample shot — run on what we sell: edges that survive it, on your TradingView charts.

View our strategies